--- title: Equivalent Wishart Ansatz in Random Matrices url: https://www.emergentmind.com/topics/equivalent-wishart-ansatz type: topic --- # Equivalent Wishart Ansatz in Random Matrices “Equivalent Wishart Ansatz” denotes a family of constructions in which a matrix-valued problem is represented, approximated, or characterized through Wishart structure. In the cited literature, the expression is used in several technically distinct senses: as a bulk small-gap universality statement for complex Wishart spectra; as the replacement of a nonsymmetric singular-value problem by a symmetric Wishart-type matrix; as an independence characterization of Wishart and matrix–Kummer laws on the positive-definite cone; as a local asymptotic equivalence between the central Wishart law and a symmetric matrix-variate normal law; as a generalized Bartlett variational posterior for deep Wishart processes; and as an effective finite-width description of Bayesian deep networks in the proportional regime through Wishart fluctuations of empirical kernels [1207.4240] [1306.2242] [1706.09718] [2104.04882] [2305.14454] [2605.29684]. Taken together, these uses suggest not a single theorem, but a recurrent methodological pattern: replacing a difficult object by an exactly Wishart, asymptotically Wishart, or Wishart-characterized surrogate. ## 1. Poissonian smallest-gap theory in Wishart spectra For the complex Wishart ensemble with \(X \in \mathbb{C}^{m\times n}\) a rectangular Ginibre matrix, \(W=X^*X\), and \(A_{mn}=X^*X/m\), the eigenvalues \(\lambda_1\le \cdots \le \lambda_n\) have joint density \[ p(\lambda_1,\ldots,\lambda_n)\propto \prod_{i0. \] The same limiting law holds for UUE after replacing \(g\) by the equilibrium density \(\Psi\), which is the precise sense in which Wishart and UUE are “equivalent” at the level of bulk smallest-gap statistics [1207.4240]. The determinantal mechanism is explicit. For \(|x-y|=\mathcal{O}(n^{-4/3})\), \[ K_n(x,x)K_n(y,y)-K_n(x,y)K_n(y,x) =\frac{1}{3}\pi^2 g(x)^4 n^4 u^2+\mathcal{O}(n), \] and the same \(u^2\)-law holds in UUE with \(g(x)\) replaced by \(\Psi(x)\). This \(2\times 2\) block asymptotics yields the intensity factor \(\int_A u^2\,du\) and therefore the exponent \(p=3\). The paper contrasts this Hermitian behavior with the complex Ginibre case, where the global smallest-gap scale is \(n^{-3/4}\) and the limiting density becomes \(\frac{4}{\Gamma(k)}x^{4k-1}e^{-x^4}\), reflecting a \(p=4\) Poissonian ansatz rather than the Wishart/UUE \(p=3\) law [1207.4240]. ## 2. Spectral reductions to Wishart-type objects A second use of the ansatz arises for nonsymmetric correlation matrices. Let \(A\in\mathbb{R}^{N\times T}\), \(B\in\mathbb{R}^{M\times T}\), \(N\le M\le T\), with Gaussian entries after decorrelation of the diagonal blocks and cross-covariance \[ \eta=\frac{1}{T}\mathbb{E}[AB^t]. \] The rectangular nonsymmetric correlation matrix is \[ X=\frac{1}{T}AB^t\in\mathbb{R}^{N\times M}, \] and the corresponding “equivalent Wishart” matrix is \[ C=\frac{1}{T^2}AB^tBA^t\in\mathbb{R}^{N\times N}. \] Its companion \(D=BA^tAB^t/T^2\) has the same nonzero eigenvalues, so the singular values of \(X\) are the square roots of the nonzero eigenvalues of \(C\). The ensemble-averaged resolvent \(G_C(z)=(1/N)\operatorname{Tr}(zI_N-C)^{-1}\) satisfies a Pastur-type self-consistent equation involving only the aspect ratios \(\kappa_N=N/T\), \(\kappa_M=M/T\), and the deterministic deformation \(\zeta=\eta\eta^t\): \[ G(z)=\frac{1}{N}\sum_{i=1}^N \frac{1}{z-\overline Y_2(z,G)-\overline Y_1(z,G)\lambda_i^{(\zeta)}}. \] In this formulation, the nonsymmetric problem is replaced by a symmetric spectral problem for \(C\), which is the operative content of the ansatz in this setting [1306.2242]. This reduction is exact enough to recover known special cases and to separate bulk deformation from outliers. When \(\eta=0\), the system reduces to the independent case, yielding the known cubic equation for the singular-value spectrum of \(AB^t/T\). When \(\zeta\) has rank one, the bulk remains that of the independent model while a single separated eigenvalue appears. More general \(\zeta\) reshapes the bulk density itself. The dependence enters only linearly through the \(\zeta Y_1\) channel, which makes the equivalent-Wishart description computationally tractable and directly tied to the deformation spectrum of \(\zeta\) [1306.2242]. A further extension appears in non-Hermitian Wishart-Laguerre ensembles. There the standard Hermitian model is generalized to a Gaussian two-matrix product, with complex eigenvalues of \(DC\) or of the associated non-Hermitian Dirac matrix \[ \mathcal{D}=\begin{bmatrix}0_N & C \\ D & 0_{N+\nu}\end{bmatrix}, \] and an interpolation parameter \(\mu\in[0,1]\) connecting Hermitian Wishart to maximally non-Hermitian regimes. The finite-\(N\) correlation functions are determined by kernels built from orthogonal or skew-orthogonal Laguerre polynomials in the complex plane, while the microscopic origin limits are governed by complex Bessel kernels. In this use, the ansatz is an extension of Wishart universality into a non-Hermitian setting rather than a reduction to a symmetric model [1104.5203]. A related but exact equivalence occurs for generalized uncorrelated Wishart matrices with deterministic zero patterns. In the complex case, if \(a_k=\theta(k-1)+c\) with \(\theta>0\) and \(c>-1\), then the eigenvalues of \[ S=(Y^{(2)})^\dagger Y^{(2)} \] have joint density \[ p(\lambda_1,\ldots,\lambda_N)\propto \prod_{i=1}^N \lambda_i^c e^{-\lambda_i}\prod_{i(p-1)/2\), \(b>(p-1)/2-a\), and \(\lambda>0\) such that \[ X\sim MK_p(a,b,\lambda I_p),\qquad Y\sim W_p(\nu,\Sigma),\quad \nu=2(a+b),\ \Sigma=(\lambda/2)I_p. \] Conversely, these laws imply independence of \(U\) and \(V\). Here the “Equivalent Wishart Ansatz” is not asymptotic: independence of a specific involutive transform is equivalent to Wishart structure with spherical scale, together with a matrix–Kummer companion law [1706.09718]. The proof passes through a functional equation for transformed log-densities, a matrix logarithmic Pexider equation on \(S_{++}^p+I_p\), and an additivity argument forcing linear trace terms. The resulting density forms consist of \(\log\det\), \(\log\det(I_p+X)\), and \(\operatorname{tr}(\lambda X)\) contributions, exactly matching the Wishart and matrix–Kummer families. The theorem is spherical: the scale is proportional to \(I_p\). The paper notes that a different transform due to Vallois allows arbitrary scale \(\Sigma\), but the converse characterization remains open [1706.09718]. An exact domain characterization appears as well for non-central Wishart distributions and Wishart processes on \(S_d^+\). Writing \(W_d(p,\Sigma,\Omega)\) for the law with Laplace transform \[ \mathbb{E}[e^{-\operatorname{tr}(UX)}] =\det(I+\Sigma U)^{-p}\exp\!\big(-\operatorname{tr}(U(I+\Sigma U)^{-1}\Omega)\big), \] the non-central Gindikin set theorem states that this law exists if and only if either \(p>(d-1)/2\) with arbitrary \(\Omega\in S_d^+\), or \(p\in\{0,1/2,\ldots,(d-2)/2\}\) with \(\operatorname{rank}(\Omega)\le 2p\). The associated Wishart SDE \[ dX_t=\sqrt{X_t}\,dW_t+dW_t^t\sqrt{X_t}+aI\,dt,\qquad X_0=x_0, \] has a global weak solution in \(S_d^+\) if and only if either \(a\ge d-1\), or \(a\in\{0,1,\ldots,d-2\}\) and \(\operatorname{rank}(x_0)\le a\). These process and distributional domains are equivalent under \(p=a/2\), and when the solution exists one has \[ X_t\sim W_d(a/2,2tI,\Omega=x_0). \] In this usage, the ansatz is valid exactly when the parameter tuple lies in the admissible non-central Gindikin domain [1607.00206]. ## 4. Local equivalence and deterministic equivalents A different formulation treats the Wishart law as locally equivalent to a Gaussian analogue. For \(W\sim W_p(n,\Sigma)\), define \[ \Delta_{n,\Sigma}(S)=(\sqrt{2n}\,\Sigma)^{-1/2}(S-n\Sigma)(\sqrt{2n}\,\Sigma)^{-1/2}, \] and let \(g_{n,\Sigma}\) denote the symmetric matrix-variate normal density with mean \(n\Sigma\) and covariance \(B_p^\top[2n(\Sigma\otimes\Sigma)]B_p\). Uniformly on the bulk set \[ B_{n,\Sigma}(\eta)=\Big\{S\in S_{p}^{++}:\max_i \big|\sqrt{2/n}\,\lambda_i(\Delta_{n,\Sigma}(S))\big|\le \eta n^{-1/3}\Big\}, \] the density ratio satisfies the explicit expansion \[ \log\frac{f_W(S;n,\Sigma)}{g_{n,\Sigma}(S)} = n^{-1/2}\Big\{\frac{\sqrt2}{3}\operatorname{tr}(\Delta^3)-\frac{p+1}{\sqrt2}\operatorname{tr}(\Delta)\Big\} + n^{-1}(\cdots) + O_{p,\eta}\!\Big(\frac{1+\max_i|\lambda_i(\Delta)|^5}{n^{3/2}}\Big). \] The paper also gives the full order-\(n^{-1}\) ratio expansion and derives \[ \operatorname{dist}(P_{n,\Sigma},Q_{n,\Sigma})\le \frac{C p^{3/2}}{\sqrt n}, \] together with the corresponding Hellinger bound. Here “Equivalent Wishart Ansatz” means local asymptotic equivalence between central Wishart and a moment-matched symmetric matrix Gaussian, with explicit correction terms [2104.04882]. This approximation is then used for a Wishart-kernel density estimator on \(S_{p}^{++}\). The estimator \[ \widehat f_{n,b}(S)=\frac1n\sum_{i=1}^n K_{1/b,bS}(X_i) \] has pointwise bias \(b\,g(S)+o(b)\), variance of order \(n^{-1}b^{-r/2}\psi(S)\) away from the boundary, and a boundary inflation factor \(b^{-|J|(p+1)/2}\) when \(|J|\) eigenvalues approach zero at rate \(b\). The local Gaussian equivalence furnishes the leading variance calculations and asymptotic normality statements [2104.04882]. Another deterministic-equivalent use arises for compound Wishart matrices \[ W_\theta=Z^T D Z, \] with \(Z\) a \(d\times n\) real Gaussian matrix of variance \(1/n\). Amplifying \(D\) to \(D\otimes I_a\) and sending \(a\to\infty\) defines free deterministic equivalent moments \[ \mu_\ell^\square(\theta)=\lim_{a\to\infty}\frac1a\mathbb{E}[\operatorname{Tr}(W_{\theta^a}^\ell)],\qquad \operatorname{Var}_\ell^\square(\theta)=\lim_{a\to\infty}\operatorname{Var}[\operatorname{Tr}(W_{\theta^a}^\ell)]. \] The resulting FDE Z-score, \[ Z_\ell^\square(\theta_0\mid\theta)=\frac{\operatorname{Tr}(W_{\theta_0}^\ell)-\mu_\ell^\square(\theta)}{\sqrt{\operatorname{Var}_\ell^\square(\theta)}}, \] converges to \(N(0,1)\) under the growth condition \[ R(D_n)=\frac{\|D_n\|}{\sqrt{\operatorname{tr}(D_n^2)}}=o(n^{1/(3\ell)}). \] The explicit formulas \[ \mu_1^\square=n m_1,\qquad \mu_2^\square=n(m_2+m_1^2)+m_2, \] \[ \operatorname{Var}_1^\square=2\lambda m_2,\qquad \operatorname{Var}_2^\square=2(4\lambda^3 m_1^2m_2+2\lambda^2m_2^2+8\lambda^2m_1m_3+4\lambda m_4) \] make the method suitable for goodness-of-fit testing of 2D ARMA models after reduction to a compound Wishart shape matrix \(D=B^TB\) [1710.09497]. ## 5. Variational Wishart parameterizations in deep probabilistic models In deep kernel processes, and specifically in the deep Wishart process (DWP), the ansatz appears as a variational posterior family over Gram matrices. A DWP alternates deterministic kernel maps \(S_l=\Phi_l(G_l)\) with Wishart sampling \[ G_{l+1}\mid G_l\sim W_N(\nu_{l+1},S_l),\qquad G_{l+1}=F_{l+1}F_{l+1}^\top, \] where the columns of \(F_{l+1}\) are i.i.d. \(N(0,S_l)\). Because standard isotropic kernels depend only on inner products and pairwise squared distances derived from \(G_l\), the DWP prior over Gram matrices is equivalent to the deep Gaussian process prior at the data locations. The variational “equivalent Wishart ansatz” is then a generalized Bartlett family, \[ G=ATT^\top A^\top \] or, in the improved form, \[ G=ATBB^\top T^\top A^\top, \] with \(A\) invertible, \(B\) invertible lower-triangular, and \(T\) lower-triangular with Gamma-distributed diagonal squares and Gaussian strictly lower entries. The additional \(B\)-mixing allows linear combinations across Bartlett columns and therefore captures posterior correlations across latent degrees of freedom at negligible extra asymptotic cost beyond the \(O(N^3)\) Gram-matrix operations already present [2305.14454]. The explicit densities follow from Jacobians for \(TT^\top\), \(AT\), and \(TB\), yielding closed-form \(q(G)\) for both the \(A\)-generalized and \(AB\)-generalized singular Wishart families. In experiments on UCI regression, the \(AB\) ansatz improved predictive performance relative to the earlier \(A\)-only posterior. Reported examples include Kin8nm at depth 4, where test log-likelihood changes from \(1.38\pm0.01\) for the DGP to \(1.40\pm0.00\) for DWP-AB, and Concrete at depth 4, where test log-likelihood changes from \(-3.12\pm0.02\) to \(-3.07\pm0.02\), with corresponding RMSE improvement from \(5.43\pm0.11\) to \(5.22\pm0.13\) [2305.14454]. A more recent use treats finite-width Bayesian deep networks in the proportional regime \(P/N_\ell=\alpha_\ell=O(1)\). There the empirical layer kernels \[ K_E^{(\ell)}=\frac{1}{\lambda N_\ell}\sigma(H^{(\ell)})\sigma(H^{(\ell)})^\top \] are modeled as Wishart after nonlinear propagation: \[ \Theta^{L-\ell}(K_E^{(\ell)})\mid K_E^{(\ell-1)}\approx \mathcal{W}_P(V^{(\ell)},N_\ell),\qquad V^{(\ell)}=\frac{1}{N_\ell}\Theta^{L-(\ell-1)}(K_E^{(\ell-1)}). \] For any contraction direction \(\bar f\in\mathbb{R}^P\), this implies independent \(\chi^2\) variables \[ Q_\ell:=N_\ell\frac{\bar f^\top \Theta^{L-\ell}(K_E^{(\ell)})\bar f}{\bar f^\top \Theta^{L-\ell+1}(K_E^{(\ell-1)})\bar f}\sim \chi^2_{N_\ell}, \] and hence a renormalized NNGP kernel \[ K_{\mathcal Q}^{(R)}=\mathcal Q\,\Theta^L(C),\qquad \mathcal Q=\prod_{\ell=1}^L q_\ell,\quad q_\ell=Q_\ell/N_\ell. \] The resulting effective action is \[ S(q)=\sum_{\ell=1}^L[q_\ell-\ln q_\ell] +\frac{\alpha}{P}\log\det(I+\beta K_{\mathcal Q}^{(R)}) +\frac{\alpha}{P}y^\top(\beta^{-1}I+K_{\mathcal Q}^{(R)})^{-1}y, \] whose saddle point determines the finite-width kernel renormalization. In CNNs the scalar \(\mathcal Q\) is replaced by a positive-definite local order parameter matrix, producing hierarchical local kernel renormalization [2605.29684]. This proportional-regime theory is approximate rather than exact, but it is non-perturbative in both \(\alpha\) and depth \(L\). The paper reports good agreement with posterior sampling for Bayesian MLPs of depth \(L\sim O(10)\) and \(P\sim O(10^3)\), together with two systematic deviations: a gradual drift at large depth and load, and a metastable regime at large \(L\alpha\) where train and test losses abruptly drop beyond a critical depth [2605.29684]. ## 6. Meanings, common structure, and limitations Across these works, “Equivalent Wishart Ansatz” has several formally different meanings. It can denote an exact law-equivalence, as in the matrix HV characterization and the non-central Gindikin/Wishart-process characterization; an exact spectral reparameterization, as in \(X=AB^t/T\mapsto C=AB^tBA^t/T^2\); an asymptotic universality statement, as in the Wishart–UUE smallest-gap correspondence; a local distributional approximation, as in the symmetric matrix-normal replacement of central Wishart; a deterministic equivalent for linear spectral statistics; or a variational/effective modeling assumption in Bayesian deep learning [1706.09718] [1607.00206] [1306.2242] [1207.4240] [2104.04882] [1710.09497] [2305.14454] [2605.29684]. Several recurring structures are stable across these settings. The underlying objects live on \(S_d^+\), \(S_{++}^p\), or Gram-matrix cones; exact or approximate equivalence is usually mediated by a factorization \(X=FF^\top\), a Bartlett-type triangularization, an affine Laplace transform, a determinantal kernel, or a Pastur-type resolvent equation. Local density dependence often enters only through low-order invariants such as \(\rho(x)^4\), \(\zeta=\eta\eta^t\), traces \(m_k=\operatorname{tr}(D^k)\), or scalar order parameters \(q_\ell\). This suggests a common principle: much of the complexity is pushed into a low-dimensional deformation of a Wishart backbone. The limitations are equally setting-specific. The smallest-gap theorem is stated only in a bulk window \(I_{\mathrm{bulk}}\), with the edge removal parameter \(\epsilon_0\) kept for technical reasons [1207.4240]. The nonsymmetric correlation model assumes Gaussianity and decorrelation of the diagonal blocks [1306.2242]. The matrix HV converse characterizes only spherical scales \(\Sigma\propto I_p\) [1706.09718]. The symmetric matrix-normal approximation is a fixed-\(p\), \(n\to\infty\) result [2104.04882]. The deep-learning formulations remain ansätze rather than proofs and are presently derived for Gaussian regression likelihoods, with the proportional-regime theory explicitly reporting systematic deviations in certain large-\(L\alpha\) regimes [2305.14454] [2605.29684]. The generalized real Wishart–Muttalib–Borodin correspondence also carries parity constraints for the explicit zonal-polynomial eigenvalue formula [2308.15001]. In that sense, the term functions less as a single doctrine than as a reusable technical motif. Whenever a matrix model can be recast in terms of a Wishart law, a Wishart-characterized transform, or a Wishart-like fluctuation principle, one gains access to a mature toolkit: Marchenko–Pastur asymptotics, Laguerre kernels, spherical functions, affine transforms, generalized Bartlett decompositions, and resolvent or large-deviation formalisms. The literature shows that this motif applies to random-matrix spectral statistics, stochastic processes on positive semidefinite cones, deterministic-equivalent testing, and Bayesian deep models alike. Source: https://www.emergentmind.com/topics/equivalent-wishart-ansatz